Key Aspects of the Research:

The study looks at how structured dialogues between several LLM agents affect the generation of high-quality ideas. The authors built a system where some agents propose ideas, others critique them, and then the first group refines the ideas based on the critique. They compared different setups: the number of critic agents, the depth (number of “critique-revision” cycles), and the diversity of the agents' “personalities” (personas).

Key Result: Using a “Generate-Critique-Revise” loop with several cycles and a diverse set of critic roles significantly boosts both the novelty and the practical feasibility of the ideas, compared with a simple single-LLM request.

The Core of the Method, Explained:

The point of the method is to stop treating the LLM as a vending machine for finished answers and start using it as a partner in an iterative process. Instead of trying to write one perfect, all-encompassing prompt, you break the task into three consecutive stages right inside one conversation:

1. Generation (Ideation): You give the LLM a clear task to produce a first version, a draft or a set of ideas. That's your starting point.

2. Critique: You ask the same LLM (or open a fresh chat for a “clean” context) to play critic. This is the most important step. You hand it its own text and ask it to find weak spots, logical gaps and risks, or to suggest improvements. The study shows it's especially effective to make the LLM critique from different perspectives by assigning it different roles (personas). For example: “Now critique this text as a marketer,” “Now as a lawyer.”

3. Revision: You give the LLM the original text plus all the critique, then issue the final command: “Taking all of these comments into account, rewrite the original text to make it stronger, more logical and more persuasive.”

You can repeat the loop 2-3 times for the best result. The study's main takeaway for users: a structured dialogue with self-critique produces far better results than a single request.

Practical Applicability Analysis:

• Direct applicability: The method works directly in any chatbot (ChatGPT, Claude, Gemini). The user can manually walk the LLM through the “Generate -> Critique -> Revise” stages. For example, first ask “Write me a sales proposal,” then “Now play a skeptical customer and find 5 weak spots in this proposal,” and finally “Great, now rewrite the proposal so it addresses all 5 objections.”

• Conceptual value: The study gives the user a key insight: quality comes from a process, not from one brilliant prompt. An LLM can not only generate but also analyze, and from different roles. That helps you see that for complex tasks you need to “make” the model look at the problem from several angles, simulating a team of experts. The finding that critic diversity (different personas) matters most for quality, while generator diversity matters most for novelty, gives real insight into how to steer a creative process.

• Adaptation potential: The method is completely universal and adapts easily to any complex task where you need not just a quick answer but high quality and careful thinking.

  • How to adapt it:
    1. Define the end goal (write an essay, draft a business plan, plan a trip).
    2. Generation: Write a simple prompt asking for a draft.
    3. Critique: Pick 2-3 key roles whose perspective on the result matters (for a business plan, an “investor” and a “marketer”; for a trip, a “parent with kids” and an “adrenaline junkie”). Ask the LLM to critique the draft from each of those roles.
    4. Revision: Collect all the critical comments and give the command to produce the final version.

A Practical Example:

### **Task: Develop a concept for a new coffee shop in a residential suburb.**

---

#### **STAGE 1: IDEA GENERATION (First prompt)**

You are a creative entrepreneur.
Your task is to develop a concept for a new coffee shop opening in a residential suburb of a big city. Target audience: families with kids, freelancers, and office workers from nearby business parks.

Propose 3 concept options. For each, specify:
1.  **Name:** Short and memorable.
2.  **Core idea (the "hook"):** What will set this coffee shop apart?
3.  **Signature menu items:** 2-3 items that reflect the concept.
4.  **Atmosphere:** What the interior will look like and what the vibe will be.

---

#### **STAGE 2: IDEA CRITIQUE (Second prompt, after you get the answer to the first)**

Great. Now switch roles.

You are an **experienced and very pragmatic investor** deciding whether to put money into these concepts. I'll give you 3 ideas that your colleague came up with.

Your task is to **critique each of the 3 concepts harshly but constructively**. For each idea, find potential weak spots, risks and gaps.

Pay attention to:
- **Financial viability:** Is the "hook" too expensive to pull off? Will it make money?
- **Audience reach:** Will the concept scare off one part of the audience while trying to attract another?
- **Operational complexity:** How hard will it be to keep this concept running day to day?

Here are the ideas to critique:
`[The user pastes the LLM's answer from stage 1 here]`

---

#### **STAGE 3: IDEA REVISION (Third prompt)**

Excellent analysis. You're the creative entrepreneur again.

You've received valuable feedback from an investor. Your task is to **pick the single most promising concept of the three and refine it** so it addresses all of the investor's objections and concerns.

Create a final, improved version of the concept. In the description, state separately **exactly how you addressed each point** of the investor's critique to make the idea stronger and more viable.

Here are the original ideas:
`[The user pastes the original answer here again]`

And here is the investor's critique:
`[The user pastes the answer from stage 2 here]`

Why It Works:

This prompt works because it puts the study's key findings into practice:

1. Task decomposition: The complex task “come up with a successful coffee shop” is split into logical steps: generation, risk analysis and synthesis of an improved solution. That prevents shallow answers.
2. Simulating a multi-agent dialogue: The user has the LLM play two different roles in turn: a creative generator and a pragmatic critic. This mimics the “Ideation-Critique” loop proposed in the paper.
3. Using personas for quality critique: The “experienced investor” role makes the model focus on specific aspects (finances, operations, risks), so the critique is targeted and useful rather than generic. This directly applies the finding that a specialized critic raises the quality and feasibility of an idea.
4. Iterative improvement: The final prompt makes the model not just rewrite the text but synthesize a new version from the original data and the critique. That taps into the LLM's more advanced “reasoning” abilities and produces a more thoughtful, balanced result.

Another Practical Example

### **Task: Plan a 5-day vacation in Lisbon for a couple.**

---

#### **STAGE 1: GENERATION (First prompt)**

Act as an experienced travel agent.
Draft a 5-day trip to Lisbon for a young couple (ages 25-30). They love history, great food and beautiful views, but don't want to spend the whole time in museums. Mid-range budget.

Present the plan as a table: Day 1, Day 2, etc. For each day, briefly describe:
- Morning (1-2 activities)
- Afternoon (1-2 activities)
- Evening (a dinner and evening option)

---

#### **STAGE 2: CRITIQUE (Second prompt)**

Thanks. Now you need to play two different roles and critique this plan.

**Role 1: "Budget traveler"**
Analyze the plan from a budget standpoint. Point out which activities or places might be too expensive. Suggest cheaper but just as interesting alternatives. Find spots where they could save on transportation or food.

**Role 2: "Fan of spontaneity and downtime"**
Analyze the plan for pace and workload. Is it overloaded? Is there time for spontaneous walks, lingering in cafes, or just doing nothing? Point out the most exhausting days or transfers.

Here is the plan to critique:
`[The user pastes the LLM's answer from stage 1 here]`

---

#### **STAGE 3: REVISION (Third prompt)**

Great work. You're the travel agent again.
You've received important feedback from clients with different priorities.

Your task is to create a final, balanced trip plan that takes both opinions into account. It should be interesting, but not too expensive and not overloaded.

In the new plan, show explicitly what changes you made to address the wishes of the "budget traveler" and the "fan of spontaneity".

Here is the original plan:
`[The user pastes the original answer here again]`

And here is the critique:
`[The user pastes the answer from stage 2 here]`

Why This Example Works.

This example applies the study's methodology effectively to a practical everyday task:

1. Modeling diverse critics: Instead of one critic, as in the first example, it uses two different personas, both relevant to the task: the “budget-minded” one and the “downtime lover.” This directly embodies the paper's key finding that diversity in the critic team (Agent Diversity) improves the final result.
2. Surfacing and resolving conflicts: The personas have potentially conflicting goals (save money vs. relax in comfort). By having the LLM first surface these conflicts (in the critique step) and then resolve them (in the revision step), the user triggers a more sophisticated reasoning process. The model has to find a compromise, which leads to a more realistic, useful plan.
3. Raising practical value (Feasibility): Critique on specific parameters (budget, pace) directly increases the plan's “feasibility,” as the study showed. The draft might just be a list of sights; the final version is a real itinerary that accounts for human factors like fatigue and budget.
4. Structured process: As in the previous example, the strict three-stage structure guides the LLM and keeps it from drifting off or giving a shallow answer. Each step builds on the last, creating a chain of reasoning and improvements.

Usefulness Score: 92